Research On Text Matching In Retrieval-based Medical Question And Answering System

With the improvement of technology,more and more attention is paid to health problems.Intelligent medical question answering system,which can quickly feedback the professional medical information,has important research and application value.Most intelligent medical question answering systems are retrieval-based question answering systems.It mainly includes two core modules: question recalling module and question re-ranking module.Question recalling module recalls top k candidate questions based on word similarity.Question re-ranking module re-ranks top k candidate questions based on semantic similarity through text matching model.Therefore,the performance of text matching model greatly affects the performance of the retrieval-based question answering system.This paper mainly focuses the text matching model in retrieval-based medical question answering system.Inspired by previous work,this paper proposes a semantic matching model based on dynamic routing and attention mechanism.Our model encodes the context information through bi-directional long short-term memory network and models the interactive matching information between two sentences by attention mechanism.Next,our model converts the resulting vectors obtained above to a fixed-length vector with dynamic routing mechanism and feeds it to the final classifier to determine the overall matching relationship.Besides,this paper constructs a large-scale medical text matching dataset for comparative experiments.The experiment shows that our model achieves the 86.66% F1 score,which is superior to the existing text matching models.In addition,the ablation experiments prove that our model can model the semantic information and matching interaction information between two sentences better.To improve the performance of text matching model,this paper introduces two different learning strategies: the fusion of Chinese multi granularity char and word information strategy and the domain adversarial transfer learning strategy.Different from English and other languages,words in Chinese are composed of chars and chars are the smallest semantic unit.In order to make full use of the word and char information,this paper proposes a learning framework which encodes the text context information from word and char channels at the same time.Then it utilizes different information fusion strategies to integrate the word and char information.Comparing with the model based on single channel words or words,word and char fusion strategy improves the F1-socre by 0.4%-2.5%.To alleviate deep learning model's poor performance problem on small-scale dataset,the adversarial learning strategy uses large-scale source domain task to improve the model performance on the small-scale target task.In this paper,the text matching model we proposed and the current mainstream text matching models are introduced into the adversarial learning strategy.The experience shows this strategy brings stable performance improvement on the two datasets.